Composable Skills + Generic Environment Adapter (CONCEPT:ORCH-1.28)¶

Overview¶

Upgrades the RLM's skill-as-SOP from raw source-mounting to composable Skill units and adds a generic environment adapter so an external environment runs through a minimal tool surface while its evaluator stays intact ("less harness"). Assimilated from predict-rlm (Trampoline-AI/predict-rlm) and the AppWorld RLM-GEPA thesis. Extends ORCH-1.12 (Predict-RLM Runtime).

How it works¶

Composable Skill (rlm/skills.py) bundles instructions + packages + modules + tools. merge_skills order-preserving-dedups packages, concatenates instructions under ## Skill: headers, and raises on module/tool name conflicts (explicit composition, no silent last-wins). PredictRLM.mount_skill_unit(skill) mounts a unit and appends its instructions.
Generic adapter (RegistryEnvironmentAdapter) exposes a small list_items / describe / call / submit surface over a callable registry; submit hands off to the host-owned evaluator — the RLM supplies only the policy (the optimizable skill), the host owns state + scoring.

Key files / API¶

Piece	Location
Skills + adapter	`rlm/skills.py` (`Skill`, `merge_skills`, `EnvironmentAdapter`, `RegistryEnvironmentAdapter`)
Mount	`rlm/predict_rlm.py` (`PredictRLM.mount_skill_unit`); `rlm/runner.py` (`run_rlm(skills=...)`)

Wiring (≤3 hops)¶

graph_orchestrate(action="rlm_run", skills=...) → run_rlm → merge_skills / mount_skill_unit (2 hops).

Research provenance¶

predict-rlm src/predict_rlm/rlm_skills.py (merge_skills, conflict detection) — verified.